What was already known
Google DeepMind’s AlphaProof and AlphaGeometry 2 systems solved four 2024 IMO problems for 28 of 42 points using expert translations into formal languages and two to three days of computation. The 2025 Gemini system received the official natural-language problem statements and ran within the 4.5-hour contest limit.
AI involvement
Google DeepMind states that Gemini Deep Think generated the proofs directly from official problem descriptions. Humans developed the system, supplied general instructions and training material, and arranged official grading.
DeepMind says the model wrote the five solutions directly from official natural-language statements within the contest limit. Human developers trained it and supplied general mathematical hints, so task-only autonomy is not established.
Importance
The result demonstrates gold-medal-level performance on known competition problems with direct natural-language input and output.
Limitations
These are solutions to competition problems, not new research theorems. IMO coordinators checked the submitted answers; they did not validate the model, training process or broader reliability. The supplied PDF contains the five solutions.
Technical verification
Official IMO coordinators graded the submitted solutions as five correct, 35/42. The IMO review concerned these answers, not the model or process.
Independent checking
The IMO president and official coordinators graded the submitted natural-language solutions as 35/42. Their review covered answers, not the model or process.
- Checker
- International Mathematical Olympiad coordinators
- Scope
- The five submitted solutions to the 2025 IMO problems, graded for a total of 35 of 42 points.
- Basis
- Reported by the linked checker
Human understanding
Few is the default when evidence for a higher level is insufficient. The coordinator quotation calls the proofs clear and precise, but the reviewed source does not document a named human reconstruction of each central argument.
Correction and publication history
Primary-source claim review for the What AI Found discovery index; non-notifying local inclusion.
Read this published revision